Your CLAUDE.md is on the weekly bill too. Most people only count one turn at a time.

The standard advice on hitting the weekly cap is: switch to Sonnet, wait for the reset, upgrade to Max 20x. All three are real options. None of them ask why the cap arrived on Wednesday. The answer for most paid users is in the config file that ships unchanged on every turn. The per-turn cost is small. The per-week cost is not.

1. The weekly cap, in one paragraph

Pro, Max 5x ($100/mo), and Max 20x ($200/mo) each have two limits. The five-hour rolling window doubled on 2026-05-06. The seven-day weekly cap got a temporary 50% bump on 2026-05-13 through 2026-07-13. After that, nobody knows. Independent measurements put Max 5x at roughly 24 Opus hours per week and Max 20x at roughly 40 Opus hours per week at the time of writing, with Sonnet hours several multiples higher. The point of this article is not the number. The point is that anything that fires on every turn pays for itself in seven-day units, not per-turn units.

2. The file

Trimmed sample for legibility. Pretend it is the first 200 tokens of a 6,000-token file (yours probably is, between stack notes, prior incidents, conventions, and skill references). One line on this snippet drives the weekly bill.

The cache-busting line is L3. The rest of the audit (vague absolutes on L9-10, missing-why on the prohibitions, the bloat on the Stack paragraph) shows up in the context-burn audit. Different scoreboard.

3. Per-turn math, then per-week math

Per-turn formula lives in src/lib/analyzer.ts (line 266 onwards):

For a 6,042-token file that is $0.906 per cache-busted session (cold input on all 30 turns) or $0.091 if the prompt cache holds. The weekly extension is one more multiplier:

4.5M input tokens does not sound large compared to the millions of tokens of code Claude actually reads. It is the floor. CLAUDE.md ships before anything else. Every Read, Bash, Grep, and subagent call piles on top of that floor. The weekly cap doesn't care which tokens are which.

4. The cache-bust amplifier

The analyzer ships one check that has more weekly impact than the other six combined. It is ten lines of TypeScript and one regex:

Three things to notice. First, the window is idx < 20 (only the first 20 lines of the file). Cache prefixes are matched from the top down; only the top mutating matters. Second, the regex matches four classes of volatile text: an ISO date 2026-05-18, the literal word today, the phrase this session, and the phrase right now. Third, the severity is high for a reason: each match flips a 6,000-token file from costing 0.6% of a Max 5x Opus week to costing 6-8% of it.

5. The weekly audit, in one terminal block

What the analyzer prints on this snippet when you tell it to use a weekly window. The UI on ccmd.dev shows the same numbers in a panel. The terminal layout is easier to screenshot.

Two numbers on the same line tell the story. The current state is in red because L3 fires the cache_bust finding. The cached state is the number you'd see one delete away. The delta is 10x. The bottom row reads in percent of your weekly Opus budget; it is the only number a CTO actually wants to see.

6. The subagent multiplier nobody counts

Every subagent invocation starts with the parent CLAUDE.md as part of its system prompt. A workflow rule in your file that says "for any non-trivial bug, spawn an Explore subagent first" means every non-trivial turn rehydrates CLAUDE.md at least twice: once for the parent, once for the subagent. Two-deep fan-outs (Explore then Plan then implement) triple it.

The weekly math, with one subagent per turn on half the turns:

• Parent turns / week: 750 (unchanged)
• Subagent turns / week: 375 (half of parent turns spawn one)
• Effective CLAUDE.md rehydrations / week: 1,125
• Cache-busted token cost: 6,042 × 1,125 = 6.8M tokens / week

On a Max 5x weekly budget that is the band where you start hitting throttles before the Friday wrap-up. On June 15, 2026, the Agent SDK credit-cost change splits subagent calls into a separate credit pool. The math gets clearer (you'll see two line items on the bill) but it does not get smaller. The same lines in CLAUDE.md keep firing.

7. Why the standard advice misses

The three pieces of advice that always come up when someone posts "hit the weekly cap on Wednesday", ranked by what they actually buy you on a typical CLAUDE.md.

Both columns are real. The point is that the first column ships today, costs nothing, and almost nobody is doing it because the per-turn math hides what the per-week math shows.

8. What to do this afternoon

1. Open ccmd.dev. Paste your CLAUDE.md into the textarea. Read the first finding. If it is a cache_bust, move that line to the bottom of the file under a heading like ## Current sprint or delete it. Cache prefix recovers next session.
2. Multiply the analyzer's per-session cost by your real weekly session count from claude.ai/usage. That number is your weekly CLAUDE.md floor. Track it.
3. Read the rest of the structural audit on the context-burn page. The cache_bust finding is the biggest single lever; the bloat, duplicate, and missing-why findings compound across 750 weekly turns.
4. If you have a shared CLAUDE.md across a team, check the drift guide. A shared file gets multiple cache_bust lines added by well-intentioned teammates; each one resets a different engineer's cache.